Jason Wei

Who he is

Jason Wei is an AI researcher whose work centres on how large language models can be made to reason and to follow instructions. His personal homepage describes him as a research scientist at Google Brain from 2020 to 2023, and says his work there helped popularize chain-of-thought prompting, instruction tuning and the study of emergence in large language models^[1]. He worked at OpenAI from 2023 to 2025 on reasoning and agents^[1], and his homepage says he currently works at Meta Superintelligence Labs^[1].

Chain-of-thought prompting

Wei is first author of “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”, submitted to arXiv on 28 January 2022. His co-authors were Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le and Denny Zhou^[2].

The idea is simple to state. Normally a model is asked a question and expected to give the answer directly. Chain-of-thought prompting instead shows the model a few worked examples in which the answer is preceded by the intermediate steps that lead to it — the “chain of thought”. Faced with a new question, the model then tends to write out similar steps before giving its final answer. Splitting a multi-step problem into smaller pieces gives the model room to carry intermediate results forward, which is hard when everything has to be produced in one jump. Because the technique changes only how the model is asked, not the model's weights, it became one of the standard ways of getting more reasoning out of a model, and it is the line of work most closely associated with Wei's name^[1].

Emergent abilities

Wei is also first author of “Emergent Abilities of Large Language Models”, submitted to arXiv on 15 June 2022, with Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Oriol Vinyals, Jeff Dean and others^[3].

That paper concerns a pattern familiar to anyone who has compared models of different sizes: some capabilities are essentially absent in smaller models and then appear, sometimes abruptly, once models pass a certain scale. Such abilities are called emergent. The practical point is about prediction: if a capability cannot be observed in small models, it is hard to extrapolate from them whether, or when, a larger model will have it. The paper helped make “emergence” a standard word in discussions of scaling^[1].

Other papers

Wei's homepage also lists “Finetuned language models are zero-shot learners” (September 2021), “Measuring short-form factuality in large language models” (October 2024) and “A simple yet challenging benchmark for browsing agents” (April 2025) among his papers^[1].

The three titles point to different problems. Instruction tuning — training a model on many tasks presented as instructions — aims at making a single model handle unfamiliar instructions without task-specific training. The factuality work belongs to the later problem of checking whether what a model says is true, at the level of short factual statements. The browsing-agent benchmark concerns systems that use the web to complete tasks, and how to test them.

Place in the field

Across these years Wei's work moved from getting reasoning out of a model by prompting it, to characterising what larger models can do at all, and then to reasoning and agents. The three organisations named on his homepage — Google Brain, OpenAI and Meta Superintelligence Labs — are among the main places where large models are built, and his papers on prompting and on scale are reference points in those discussions^[1].